GPT-5.4 Pro Solves Longstanding Erdős Math Problem

On April 15, 2026, the landscape of theoretical mathematics and artificial intelligence underwent a seismic shift. OpenAI’s GPT-5.4 Pro, the latest iteration of its flagship frontier model, reportedly solved a longstanding open problem in Erdős discrepancy theory in just 80 minutes. The breakthrough, which focuses on Erdős Problem #1196, was not merely a feat of computational power but a display of genuine mathematical creativity. The solution was validated by Terence Tao, a Fields Medalist and one of the world’s leading mathematicians, who described the model’s contribution as “a meaningful advancement in the anatomy of integers” that transcends the specific problem itself.

The achievement marks a definitive transition for Large Language Models (LLMs) from “stochastic parrots” capable of literature retrieval to autonomous intellectual agents capable of novel discovery. By bridging the gap between informal reasoning and formal mathematical proof, GPT-5.4 Pro has effectively demonstrated that the “System 2” reasoning—deliberative, logical, and self-correcting—is no longer the sole domain of the human mind.

The Solving of Erdős Problem #1196: A New Frontier in Combinatorics

The problem solved by GPT-5.4 Pro involves discrepancy theory, a branch of combinatorics and number theory that investigates the inevitable irregularities in the distribution of sequences. Specifically, Erdős Problem #1196 addresses the behavior of partial sums in sequences of integers and their relationship to arithmetic progressions. While the general Erdős Discrepancy Problem was solved by Terence Tao in 2015, several specific conjectures regarding the “anatomy of integers” remained stubbornly open for decades.

According to reports from the mathematical community, GPT-5.4 Pro did not just provide a raw answer; it generated a comprehensive LaTeX research paper in under 30 minutes following its initial 80-minute “thinking” phase. The model’s breakthrough relied on an unexpected connection between Markov process theory and the distribution of prime factors—a link that human mathematicians had theorized but never successfully formalized. Tao noted that the model identified a “novel piecewise eigenvector construction” that simplified the complex 13-page approach previously attempted by human researchers.

This success was facilitated by a multi-layered verification pipeline:

• Informal Reasoning: GPT-5.4 Pro acted as the “creative brainstormer,” proposing the core mathematical strategy.
• Formal Verification: The model’s output was translated and verified using the Lean proof assistant, ensuring that every logical step was mathematically sound.
• Human Oversight: Experts like Tao and Kevin Barreto provided the high-level framing and final validation of the proof’s significance.

The “Thinking” Variant and the Power of Test-Time Compute

The primary driver behind this breakthrough is the model’s “Thinking” variant. Unlike previous versions that generated tokens in a single forward pass, GPT-5.4 Pro utilizes test-time compute (also known as inference-time scaling). This architecture allows the model to “pause” and allocate additional computational resources to evaluate its own internal reasoning steps before finalizing a response.

In technical terms, this represents a shift from System 1 thinking (fast, intuitive, and prone to error) to System 2 thinking (slow, logical, and deliberative). By internally iterating on complex logical chains, GPT-5.4 Pro can identify contradictions in its own arguments and self-correct—a process that was essential for solving a problem as abstract as the Erdős discrepancy. This deliberative process allows the model to handle long-horizon reasoning, where the complexity of the task requires maintaining context across thousands of logical operations without succumbing to the “hallucinations” that plagued earlier LLMs.

Test-Time Compute vs. Training Compute

For years, the industry focused on scaling laws related to training data and model size. However, GPT-5.4 Pro proves that scaling inference compute is equally, if not more, critical for high-stakes problem solving. By allowing the model to “think” longer—effectively searching through a larger space of possible solutions—OpenAI has unlocked a level of accuracy that was previously unattainable, even with models trained on larger datasets.

OSWorld-Verified: Surpassing the Human Baseline in Computer Use

While the mathematical breakthrough captured the headlines of the academic world, another record was shattered in the realm of practical automation. GPT-5.4 Pro set a new record on the OSWorld-Verified benchmark, scoring an unprecedented 75.0%. This represents a staggering 27.7% increase over its predecessor, GPT-5.2, and, more importantly, it marks the first time an AI model has surpassed the human expert baseline of 72.4% on this specific evaluation.

The OSWorld-Verified benchmark measures a model’s ability to act as an autonomous agent within a standard desktop environment. This includes:

• Navigating complex terminal interfaces and file systems.
• Using web browsers to conduct research and fill out multi-step forms.
• Operating desktop software via mouse and keyboard commands based on visual screenshots.
• Coordinating tasks across multiple applications simultaneously.

The leap to 75.0% suggests that GPT-5.4 Pro can now perform real-world knowledge work with a reliability that rivals, or exceeds, that of a human professional. This is not merely about following simple instructions; it is about high-level planning. For the Erdős problem, the model likely used its autonomous computer-use capabilities to navigate mathematical databases, run local simulations to test its hypotheses, and manage the Lean verification environment—all without constant human prompting.

Architecture of the “Unified” Model: Coding, Reasoning, and Agency

One of the most significant aspects of GPT-5.4 Pro is its unified architecture. In previous years, OpenAI often split its research into specialized models, such as GPT-5.3-Codex for programming or the o-series for reasoning. GPT-5.4 Pro integrates these capabilities into a single system, supported by a 1 million token context window.

This unification allows for a “build-run-verify-fix” loop that is native to the model. When tasked with a research problem, GPT-5.4 Pro can:

1. Plan: Use its “Thinking” variant to outline a multi-step research strategy.
2. Execute: Write the necessary Python or Lean code using its frontier coding skills.
3. Act: Use its computer-use capabilities to run the code in a terminal or browser.
4. Verify: Analyze the output and, if errors are detected, use its reasoning engine to debug and re-run the task.

This loop is further optimized by a new “Tool Search” feature, which OpenAI claims reduces token usage by 47% on tool-heavy tasks. Instead of loading every available tool definition into the context window at once, the model searches for and loads only the relevant tools on demand, significantly increasing the efficiency of long-horizon trajectories.

Implications for the Future of Science and Professional Work

The successful resolution of an Erdős problem by GPT-5.4 Pro signals the “industrialization” of mathematical and scientific research. As Terence Tao observed, the real win is not just the solution itself but the speed at which AI can draft, revise, and verify mathematical texts. Work that previously took months of grueling effort from PhD-level researchers can now be “ballparked” in under two hours.

The Rise of the Digital Research Assistant

We are entering an era where AI models act as digital lab assistants. In the fields of financial modeling, legal analysis, and medical diagnostics—where the cost of an error is immense—the GPT-5.4 Pro variant is designed to minimize risk. OpenAI reports that the model produces 33% fewer factual errors than GPT-5.2, making it a viable tool for professional workflows that demand absolute precision.

A Paradigm Shift in Knowledge Work

The 83.0% score on the GDPval benchmark (measuring performance across 44 real-world occupations) confirms that GPT-5.4 Pro is increasingly capable of handling the “boring” but complex middle-management tasks of the modern economy. From synthesizing multi-source research reports to managing intricate spreadsheets, the model is evolving from a chat interface into a fully functional operating system for intelligence.

Conclusion: The Dawn of the Reasoning AI Era

The events of April 15, 2026, will likely be remembered as the moment the “Stochastic Parrot” argument finally died. GPT-5.4 Pro has proven that through the scaling of test-time compute and the integration of autonomous agency, AI can contribute original, verified knowledge to the most rigorous fields of human study. While mathematicians like Terence Tao caution that we are still in the early stages of this “AI-assisted research” era, the trajectory is clear.

By solving a problem that resisted the brightest human minds for decades, GPT-5.4 Pro has not just solved a math problem—it has solved the problem of AI reliability. As these models continue to scale their ability to think, act, and verify, the boundary between human and machine intelligence will continue to blur, ushering in a new age of accelerated scientific and technological progress.